字号·· | 护眼
纽约时报

如果没有人工智能泡沫会怎样?What if There Is No A.I. Bubble?

点「原文对照」整页切到原文,或双击某段只看那段的原文。

Anthropic 即将上市。这可能成为有史以来规模最大的首次公开募股(IPO),并带来一笔意外之财,极大地加速该公司在人工智能行业称霸的计划。Anthropic 首席执行官达里奥·阿莫代伊、他的公司以及硅谷的许多人都感到欣喜若狂。但华尔街却如坐针毡。

Anthropic is about to go public. It could be the biggest initial public offering on record and bring in a windfall of cash that greatly accelerates the company’s plans to dominate the artificial intelligence industry. Anthropic’s chief executive, Dario Amodei, his company and many people in Silicon Valley feel jubilant. But Wall Street is on tenterhooks.

投资者担心,科技公司无法赚取足够的利润来覆盖其在人工智能数据中心上的支出。只需看看中期选举后科技股的抛售潮,就能看出这种动摇的信心。

Investors fear that tech companies cannot earn enough to cover what they are spending on data centers for A.I. One need look no further than the tech stock sell-off after the midterm elections for signs of this shaken confidence.

美国科技巨头在 2025 年至 2030 年间可能花费 5.3 万亿美元。预期回报充满不确定性。支付这笔账单到本世纪末可能需要每年 2 万亿美元的收入。那是该行业目前收入的 10 倍,也超过全球每年在软件上的总支出。最简单的结论是:收入无法达到那个高度,崩盘不可避免。

America’s tech titans might spend $5.3 trillion between 2025 and 2030. The promised returns are uncertain. Paying that bill might require $2 trillion in annual revenue by the end of the decade. That is 10 times what the industry brings in today and more than the world spends on software each year. The simplest conclusion is that revenues can’t get there and a bust is inevitable.

这种观点可能是错的。60 年来,科技行业出售的是工具:薪资处理、预算电子表格、演示软件。那些工具帮助了员工——并没有取代他们。账单由办公室的 IT 部门买单。

That view might be wrong. For 60 years, the tech industry sold tools: payroll processing, spreadsheets for budgets and software for presentations. Those tools helped employees — they didn’t replace them. The bill went to the office’s I.T. department.

人工智能不同。它越来越多地能够完成此前由分析师、会计师和律师主导的工作。它竞争的是最大的一笔企业开支:工资单。在美国,白领工资总额每年约为 8 万亿至 10 万亿美元。那是每一位理赔专员、市场协调员、呼叫中心员工、软件工程师以及其他数百万人的薪水总和。在全球范围内,这个数字大得多。

A.I. is different. It can increasingly do more of the work previously helmed by analysts, accountants and attorneys. It competes for the largest corporate expense of all: the wage bill. In America, white-collar wages total roughly $8 trillion to $10 trillion a year. That is the salary of every claims adjuster, marketing coordinator, call-center worker, software engineer and millions of others. Globally, that number is much larger.

如果企业将资金投入算力而非薪资,人工智能投资热潮或许能实现自我回报。(微软可能正在押注这一点,它向6万名员工提供买断方案,试图将更多资金转向人工智能。)经济或许能避免泡沫破裂。但代价将由工人承担,他们不得不通过招聘放缓、薪资增长疲软,最终面临失业来消化这种冲击。你认为这种权衡是否值得见仁见智;但社会以及那些有能力应对这一未来的人,规划如何应对至关重要。

An A.I. investment boom could pay for itself if companies invest in computing instead of salaries. (That’s already the bet Microsoft may be making as it offers buyouts to 60,000 employees, in an effort to move more money into A.I.) The economy might avoid a burst bubble. The cost of doing so would fall to workers, who would have to absorb the disruption through slower hiring, tepid wage growth and, ultimately, lost jobs. Whether you believe this trade-off is worth it is up to you; it’s nonetheless important that society, and those in a position to do something about this future, plan for how to respond.

老板们认为人工智能可能取代员工的首个迹象,将是技术支出溢出企业信息技术预算。这将表明企业视人工智能为不仅仅是传统软件。

The first sign that bosses reckon A.I. could supplant employees would be if spending on the technology spills out of company information technology budgets. This would indicate that the companies see A.I. as more than traditional software.

一项针对大型企业1300名高管的最新调查显示,这种情况已经出现。这些企业的人工智能预算在短短一年内翻了两番,现占其营收的3.3%。最引人注目的发现是:每花在人工智能上的1美元中,只有15美分归属于IT部门。其余资金分散至市场、产品和销售部门——正是那些雇佣员工的部门。除非这些部门的预算增加,否则人工智能支出将与工资成本展开竞争。

A recent survey of 1,300 executives in large companies suggests it already is. A.I. budgets in these businesses have tripled in just a year, now representing 3.3 percent of their revenues. The extraordinary finding: Only 15 cents of each dollar spent on A.I. lie in the I.T. department. The rest is spread to marketing, product and sales — right to the departments that hire people. Unless the budgets for those departments grow, A.I. spending will compete with the costs to cover wages.

另一项调查显示,尽管仅有16%的美国老板认为人工智能帮助他们销售更多,但60%的人预计到2029年能获得此类回报。值得注意的是,该调查中48%的美国受访者预期人工智能将在未来三年内减少其公司的就业岗位,而仅有14%预期招聘会增加。

Another survey shows that while only 16 percent of American bosses believe that A.I. has helped them sell more, 60 percent expect to reap such rewards by 2029. Tellingly, 48 percent of the survey’s American respondents expected A.I. to reduce employment in their companies in the next three years, while only 14 percent expected hiring to increase.

这与我听到的情况相符。美国和欧洲数十位高管向我抱怨,人工智能回报太慢,技术比预期更复杂。但他们对最终取得成果的信心正在增强。没人打算放缓脚步,所有人都表示明年将增加在人工智能上的投入。

This matches what I hear. Dozens of executives in the United States and Europe complain to me that A.I. is paying off too slowly and that the technology has turned out to be more complex than expected. But the confidence that they will get results is growing. Nobody plans to slow down, and all say they will spend more on A.I. next year.

如果他们的直觉正确,这将迅速累积。Anthropic 的经济学家最近进行了一项研究,试图确定人工智能可能带来的经济未来。其中一种情景估计,到 2030 年,人工智能可能为美国国内生产总值增加 2.8 万亿美元 —— 足以支付这些技术目前的投资。

If their intuitions are correct, this will add up quickly. The economists at Anthropic ran a recent study trying to determine possible futures of the economy resulting from A.I. One scenario estimates A.I. might add $2.8 trillion to the American gross domestic product by 2030 — more than enough to pay for the investments being made in these technologies.

不过,这笔国民收入的增量分配不会均匀。在同一情景下,Anthropic 估计知识工人失业的可能性可能比现在高出近一半,而他们的收入保持不变。在整个经济中,作为工资支付的国民收入份额可能下降,幅度约等于 1970 年后那 50 年下降的总和,当时工会在美国经济中仍是一股力量。那次长期滑落 —— 回报从工人转移到所有者 —— 构成了当今许多政治局势的基础。新的滑落可能在不到五年内发生。

That boost to national income will not be shared evenly, though. In that same scenario, Anthropic estimates knowledge workers might be nearly half-again as likely to be out of a job as they are today while their earnings remain the same. Across the economy, the share of national income paid as wages might fall by about as much as it did in the 50 years of decline that followed 1970, back when labor unions were still a force in the American economy. That long slide, of rewards shifting from workers to owners, sits under much of today’s politics. The new slide might take less than five years.

自动化在做大蛋糕的同时让部分人挨饿,这并非首次。1842 年,一位恼火的德国父亲把儿子送到英国曼彻斯特郊外家族棉纺厂工作,曼彻斯特是世界上第一个工业化城市。他希望这能治愈年轻人的激进观点。结果并没有。

This wouldn’t be the first time automation has grown the pie while leaving some hungry. In 1842, an exasperated German father sent his son to work at the family’s cotton mill on the outskirts of Manchester, England, the world’s first industrialized city. He hoped this would cure the young man of his radical opinions. It did not.

儿子弗里德里希·恩格斯深入贫民窟,记录下所见所闻。他的记述成为了英国历史上一段长达50年时期的缩影,这段时期始于1790年,当时人均产出上升,工资却几乎停滞不前;这段时间被称为“恩格斯停顿”。并未出现经济崩溃。磨坊主们积累了世代财富,但过了50年工资才开始追上来。

The son, Friedrich Engels, explored the slums and chronicled what he observed. His accounts came to exemplify the 50-year period in British history, starting in 1790, when output per worker climbed but wages barely moved; the time became known as Engels’ pause. There was no bust. Mill owners made generational fortunes, but it was 50 years before wages began to catch up.

这不仅仅是遥远的历史。一项关于2000年至2016年荷兰自动化的研究显示,在实现自动化的公司工作的员工收入更低,且更难找到新工作。

This isn’t just distant history. A study of automation in the Netherlands between 2000 and 2016 showed that workers in companies that automated earned less and were more likely to struggle to find new jobs.

我们不应排除人工智能将创造新工作的可能性。美国约五分之三的人从事的是1940年以来新创造的职业。1994年我是一名网站管理员——四年前我上大学时,没人想过会有这个职业。人工智能将创造新工作:也许在网络安全领域,也许在验证人工智能生成内容的领域。但这些工作可能要几年后才会出现,数量较少,且距离最先感受到冲击的人群有数百英里之遥。

We should not rule out the new jobs A.I. will create. About three-fifths of people in the United States are in kinds of jobs created since 1940. I was a webmaster in 1994 — a job that no one had even thought of when I started college four years earlier. A.I. will create new jobs: perhaps in cybersecurity, perhaps in verifying the work produced by A.I. itself. But those jobs may arrive years later, in smaller numbers, and hundreds of miles from the people who feel the effects first.

约翰·梅纳德·凯恩斯早已预见到这一点。1930年,他警告过技术性失业,认为这只是“暂时的不适应阶段”。从长远看,他终将被证明是对的。他深知长远的麻烦。但企业老板们是在短期内做决策——而且他们似乎一心要证明人工智能能盈利。指望市场自行运作迅速弥合这一鸿沟,极不可能。

John Maynard Keynes saw this coming. In 1930, he warned of technological unemployment, seeing it as only a “temporary phase of maladjustment.”He will, eventually, be right. He knew the trouble with the long run. But corporate bosses are making their decisions in the short run — and they seem intent on showing that they can make A.I. pay. It is highly unlikely that the market, left to its own devices, will close this gap quickly.

人工智能泡沫破裂会造成伤害。破裂通常喧嚣、狂乱,难以被忽视。我们可以想象企业巨头们大步走进闭门峰会,电视上专家们滔滔不绝地发表议论。

An A.I. bust would hurt. Busts are usually noisy, frantic affairs that are hard to miss. We can picture the corporate titans speed-walking into closed-door summits and the parades of talking heads opining on television.

但一场靠榨取美国工资总额来自我买单的繁荣将是无声的。它表现为从未发布的招聘广告、无法踏上晋升第一阶梯的年轻毕业生,以及领工资者收入缓慢却切实地下降。

But a boom that paid for itself by drawing on America’s wage bill would be quiet. It would be job ads never posted, young graduates who can’t reach the first rung of advancement and wages that fall slowly but perceptibly for those who earn them.

如果今天的人工智能政治局势糟糕,明天将会恶化得多。这种情景值得提前准备,而这正是华尔街未曾关注的。

If the politics of A.I. are bad today, they will get much worse tomorrow. That scenario merits preparation, and it is one Wall Street isn’t listening for.

阿齐姆·阿扎尔是一名投资人,也是 Exponential View 的创始人。

Azeem Azhar is an investor and the founder of Exponential View.

阅读媒体原文 ↗